Step 3.7 Flash

StepFun's Step 3.7 Flash, released at the end of May 2026 under an Apache 2.0 license, is drawing attention as an open-weights multimodal model that pairs strong claimed benchmark performance with low active compute, positioning the Shanghai-based lab as a serious competitor in inference efficiency, local deployment, a The model is a sparse mixture-of-experts vision-language system with a 196B-parameter language backbone, a 1.8B-parameter vision encoder, and roughly 11B activated parameters per token, making the active compute bill more relevant than headline size for developers and startups. According to StepFun's Hugging Face model

StepFun released Step 3.7 Flash on May 29, 2026, as a multimodal Mixture-of-Experts vision-language model targeting agentic use cases, succeeding Step 3.5 Flash. The model is described as a 198B-parameter sparse MoE system combining a 196B-parameter language backbone with a 1.8B-parameter vision encoder, adding native Coverage of the launch emphasizes the model's agentic framing, with native multimodal capability and tool-use enhancements highlighted as the key deltas from Step 3.5 Flash. The article treats Step 3.7 Flash as StepFun's entry into the efficiency-focused multimodal MoE category, aimed at developer workflows that requir

Step 3.7 Flash is a high-efficiency, production-level agent model released and open-sourced by the Shanghai-based foundational AI lab StepFun on May 29, 2026 under an Apache 2.0 license. It belongs to StepFun's Flash series, succeeding Step 3.5 Flash, and was explicitly designed to balance speed, cost, reliable executi The model uses a Sparse Mixture-of-Experts architecture with a 196B-parameter language backbone plus a 1.8B-parameter vision encoder, activating only about 11B parameters per token and reaching up to 400 tokens/s generation speed. It natively supports multimodal understanding and execution, internet and visual search,